libclc/r600: Use target specific builtins to implement rsqrt and native_rsqrt

[jvesely]

Summary

Fixes OCL CTS rsqrt and half_rsqrt (1 thread, scalaer) tests on AMD Turks.

[awatry]

Looks good to me from a code standpoint.

The test fails CTS before and after on my BARTS (6850), but that's when running libclc master with LLVM 9.0, so there might be a precision issue that's fixed in newer code. It doesn't make the issue worse, at least.

[jvesely]

In D74016#1865339, @awatry wrote:

Looks good to me from a code standpoint.

The test fails CTS before and after on my BARTS (6850), but that's when running libclc master with LLVM 9.0, so there might be a precision issue that's fixed in newer code. It doesn't make the issue worse, at least.

thanks.
that's surprising. does it fail for scalar only? do you have the asm dump?
I thought Barts and Turks were largely identical when it came to compute pipeline
This should really produce just single instruction so newer LLVM is unlikely to fix it.

libclc/r600/lib/SOURCES

 math/fmax.cl
 math/fmin.cl
+math/native_rsqrt.cl
+math/rsqrt.cl
 synchronization/barrier.cl
 workitem/get_global_offset.cl
 workitem/get_group_id.cl

libclc/r600/lib/math/native_rsqrt.cl

+#include <clc/clc.h>
+
+#include "../../../generic/lib/clcmacro.h"
+
+_CLC_OVERLOAD _CLC_DEF float native_rsqrt(float x)
+{
+    return __builtin_r600_recipsqrt_ieeef(x);
+}
+
+_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, float, native_rsqrt, float);

libclc/r600/lib/math/rsqrt.cl

+#include <clc/clc.h>
+
+#include "../../../generic/lib/clcmacro.h"
+
+_CLC_OVERLOAD _CLC_DEF float rsqrt(float x)
+{
+    return __builtin_r600_recipsqrt_ieeef(x);
+}
+
+_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, float, rsqrt, float);
+
+#ifdef cl_khr_fp64
+
+#pragma OPENCL EXTENSION cl_khr_fp64 : enable
+
+_CLC_OVERLOAD _CLC_DEF double rsqrt(double x)
+{
+    return __builtin_r600_recipsqrt_ieee(x);
+}
+
+_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, double, rsqrt, double);
+
+#endif